Battery simulator health diagnosis method based on gating input weight self-adaption

Abstract

The invention provides a battery simulator health diagnosis method based on gating input weight self-adaption. The method comprises the following steps: decomposing a current signal into a plurality of sub-signals by adopting a correlation noise-adding reconstruction method, and reconstructing the sub-signals by adopting a correlation coefficient method to obtain a denoised current signal; inputting the de-noised current signal into a time sequence KAN module to extract a time sequence characteristic of the signal; inputting the time sequence features into a multi-dimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features; performing feature fusion on the time sequence features, the multi-scale features and the multi-dimensional features by using a residual path attention mechanism to obtain evaluation features; distance measurement between input evaluation features and deep embedding measurement network learning features is carried out, and a discriminative feature space is obtained to judge the state of a sample. The invention provides a battery simulator performance evaluation method. According to the method, the performance of a battery simulator is evaluated to prevent circuit precision reduction caused by poor performance of the battery simulator.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence circuit fault diagnosis, and specifically relates to a battery simulator health diagnosis method based on gated input weight adaptation. Background Art

[0002] The battery simulator is a key tool that is widely used in the development, testing, and verification of battery-powered systems and related devices. It can accurately simulate the charge and discharge characteristics of different types of batteries, including voltage, capacity, and internal resistance, helping engineers to conduct a comprehensive performance evaluation of the device in a laboratory environment. This simulation enables developers to test the performance of the device under different battery states and quickly verify the power management system, such as the charger, discharge control circuit, or power management IC. In addition, the battery simulator can simulate battery failure or aging conditions (such as overcharge, overdischarge, short circuit, etc.), thereby helping developers verify the response and safety of the device under extreme or abnormal conditions. In the R&D stage, the battery simulator can reduce the dependence on real batteries, reduce testing costs, reduce resource waste, and accelerate product iteration. It also supports integration with complex systems and enhances the flexibility of testing. It is suitable for fields such as electric vehicles, renewable energy systems, and consumer electronics.

[0003] Currently, there is no method that can quickly and accurately evaluate the performance of a battery simulator. Summary of the invention

[0004] The purpose of the present invention is a method for evaluating the performance of a battery simulator, which can accurately and quickly predict the performance of the battery simulator based on a current signal.

[0005] In order to solve the above technical problems, the present invention proposes a battery simulator health diagnosis method based on gate input weight adaptation, comprising: The current signal is decomposed into multiple sub-signals using the correlation noise reconstruction method, and the sub-signals are reconstructed using the correlation coefficient method to obtain the denoised current signal; The denoised current signal is input into the gated weight adaptive module to extract the timing characteristics of the signal; After using the KAN network to extract nonlinear features from the denoised current signal, the input gated input weight adaptive module is used to extract the timing features; Input the time series features into the multi-dimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features; The time series features, multi-scale features and multi-dimensional features are fused using the residual path attention mechanism to obtain the evaluation features; Through the input evaluation features, the deep embedding metric network learns the distance metric between features and obtains a discriminative feature space to judge the state of the sample.

[0006] Optionally, the method of using a correlation noise-adding reconstruction method to decompose the current signal into a plurality of sub-signals, and using correlation coefficients to screen and reconstruct the sub-signals to obtain a denoised current signal includes: The correlated noise reconstruction method is used to add specific noise to the collected current signal and then decompose it into sub-signals that can display dimensional feature information. The correlation coefficient method is used to obtain the correlation coefficient between the sub-signal and the original signal. The sub-signals with lower correlation coefficient values ​​are removed, and the weighted sum of the remaining sub-signals is calculated to obtain the denoised current signal.

[0007] Optionally, after extracting nonlinear features from the denoised current signal using the KAN network, inputting a gated input weight adaptive module to extract timing features includes: The KAN network is used to extract high-dimensional nonlinear features in the denoised current signal. The KAN network can enhance the expression ability, use memory units to store short-term memory and capture the dynamic changes of data in time series, and combine the gating mechanism to dynamically control the information flow to obtain timing characteristics.

[0008] Optionally, the step of inputting the time series features into a multi-dimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features includes: The temporal features are input into the multi-dimensional feature module, which mainly includes three channels. The first channel uses a multi-scale network to extract multi-scale features. The second channel uses spatially separable convolution to obtain more delicate temporal features. The third channel uses maximum pooling to retain large-scale features of high-dimensional variation information. The features of the second and third channels are fused to obtain multi-dimensional features.

[0009] Optionally, the time series features, multi-scale features and multi-dimensional features are fused using a residual path attention mechanism to obtain evaluation features, including: The temporal features, multi-scale features and multi-dimensional features are concatenated to generate a path feature map. The original features are compressed in a specific dimensional direction to obtain the representative values ​​of the time and channel dimensions of each path. The attention scores calculated by the maximum pool and the average pool are then added together to obtain the path vector. The path feature map and the path activation vector are multiplied to obtain the evaluation feature.

[0010] Optionally, the deep embedding metric network learns the distance metric between the input evaluation features to obtain a discriminative feature space to judge the state of the sample, including: The evaluation features are input into the metric classifier to create a discriminant space and the network is trained through ternary metric error and mean square error to obtain the classification threshold. The loss value of the sample in the discriminant space and the classification threshold are used to judge the performance of the circuit.

[0011] The present application proposes a battery simulator health diagnosis method based on gated input weight adaptation. The method includes proposing a correlation noise reconstruction method to decompose the current signal into multiple sub-signals, and reconstructing the sub-signals using the correlation coefficient method to obtain a denoised current signal; using the KAN network to extract nonlinear features from the denoised current signal, and then inputting the gated input weight adaptation module to extract timing features; inputting the timing features into the multidimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features; using the residual path attention mechanism to fuse the timing features, multi-scale features and multi-dimensional features to obtain evaluation features; through the input evaluation features, the distance measurement between the features is deeply embedded in the metric network to obtain a discriminative feature space to judge the state of the sample. The present application provides a battery simulator performance evaluation method, which evaluates the performance of the battery simulator to prevent the poor performance of the battery simulator from causing a decrease in circuit accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0013] Figure 1 A schematic flow chart of a battery simulator health diagnosis method based on gated input weight adaptation provided in an embodiment of the present application; Figure 2 A schematic diagram of a multi-scale module of channel two provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0015] like Figure 1 Show, Figure 1 The flowchart of the battery simulator health diagnosis method based on gated input weight adaptation provided in the embodiment of the present application specifically includes five contents.

[0016] S11: A correlation noise reconstruction method is proposed to decompose the current signal into multiple sub-signals, and the correlation coefficient reconstruction method is used to reconstruct the sub-signals to obtain a denoised current signal.

[0017] It should be noted that the collected current data is usually affected by the external environment or equipment status, and has non-stationary and nonlinear characteristics. Traditional linear analysis methods are difficult to effectively extract features. The noise adaptive decomposition method can adaptively decompose the signal, remove external noise and capture the complex changes therein.

[0018] The steps of transforming and decomposing the current signal x(t) using the correlation noise reconstruction method are as follows: Step 1: Add a signal-to-noise ratio of Standard normal white noise , and obtain I noisy current signal ; Step 2: Take I noisy current signals as input, as shown in formula (2), use EMD method to decompose and obtain I component results, average the I component results to obtain sub-signal S1 and residual term res1(t); Step 3: Add standard normal white noise to the residual term, and then continue to decompose according to step 2 until the residual signal is a monotonic function or a constant. Then, N sub-signals S and the final residual signal are obtained. The noise adaptive decomposition method decomposition formula of the original signal is shown in formula (3): Where N is the final number of decompositions; Step 4: Use the correlation coefficient method to calculate the correlation coefficients of the N sub-signals and the original signal respectively. The correlation coefficient Ri corresponding to the i-th sub-signal Si is calculated by formula (4): Step 5: Normalize the obtained correlation coefficient to obtain ri, remove the sub-signals below the correlation threshold Z, and then use the normalized coefficient as the weight to reconstruct the signal as shown in formula (5) to obtain the denoised current signal s (u):

[0019] Through the above method, the noise adaptive decomposition method solves the modal aliasing problem when processing signals by introducing adaptive noise, while maintaining the adaptive decomposition ability of the current signal. Then, the correlation coefficient of each modal component is calculated by the correlation coefficient weighted method, and different weights are assigned to them by the correlation coefficient. Then, in the reconstruction process, different weights are used to reduce the impact of invalid or noisy modes on training.

[0020] Based on the above discussion, in an optional embodiment of the present application, a noise adaptive decomposition method is proposed to decompose the current signal into multiple sub-signals, and a correlation coefficient reconstruction method is used to reconstruct the sub-signals to obtain a denoised current signal. It may include: In an optional embodiment of S11, when the current signal x(t) is transformed and decomposed, I=8 Gaussian white noises are added respectively, and the signal-to-noise ratio of the added noise is =23dB, set the correlation threshold Z=0.12.

[0021] S12: After using the KAN network to extract nonlinear features from the denoised current signal, the gated input weight adaptive module is input to extract timing features.

[0022] It should be noted that the combination of LSTM units and KANs enables the use of KAN's learnable activation function to capture complex nonlinearities, and also enables the use of LSTM unit architectures to maintain long-term memory of past events. This combination provides superior modeling capabilities for tasks involving complex sequential data.

[0023] The denoised current signal (x, t) is input into the KAN module, and the KAN network is used to extract the high-dimensional nonlinear features in the denoised current signal. As shown below: in, Represents a KAN layer. The two Between' The ' symbol indicates that two KAN layers are stacked together, (l, i) represents the i-th neuron in the l-th layer, and x l,i represents the activation value of the (l, i) neuron. Between layer l and layer l+1, there are n l *n l+1 The activation function connecting (l, i) and (l + 1, j) is expressed as shown in formula (7):

[0024] Pre-activate Expressed as , after activation Expressed as = The activation value of a neuron is the sum of all activations after all inputs: The activation function of the l-th layer KAN network is expressed in matrix form as shown in formula (9):

[0025] Use memory units combined with KAN networks to preserve short-term memory and capture dynamic changes in data in time series, and calculate the output of memory units It is expressed by formula (10): Where W is the weight vector, is the hidden state of the previous time state, and Weight the importance of past values ​​and recent input separately.

[0026] Combined with the gating mechanism to process the above nonlinear features, the forget gate will choose to forget the features that are irrelevant to the current fault diagnosis. The output f of the forget gate at time t t From formula (11), we can get: in Indicates that in the forget gate f t In the state, for input x t The weight parameter, W hf Indicates the hidden state h of the previous moment under the forget gate f state t-1 The weight parameter, b f Represents the bias vector in the state of forget gate f.

[0027] The input gate processes the one-dimensional current time series data to determine the amount of new information retained in the memory cell. The output i of the input gate t As shown in formula (12): where w xi Indicates that for input x in the state of input gate i t The weight parameter of i represents the bias vector in state i, x t represents the input at the current time t, h t-1 Indicates the hidden state of the previous moment: W hi Indicates the hidden state h of the previous moment under the input gate hi state t-1 The weight parameter of .

[0028] Memory cells are used to store and transmit long-term memory information, and determine the current characteristic information that should be remembered, forgotten, or updated. Represents the candidate vector generated at time t, which is determined by the hidden information of the previous time step and the current input. The calculation formula is shown in formula (13): Among them, W xc Indicates that in state C, for input x t The weight parameter, W hc Indicates the hidden state h of the previous moment in state C t-1 The weight parameter, bc represents the bias vector in state C, the state vector C at time t t From formula (14), we can get: Among them C t-1 Represents the memory cell of the previous time step.

[0029] The output gate determines whether to retain the current characteristic information and its importance. The output of the output gate O t Calculated by formula (15):

[0030] According to the output gate output and the state vector C at time t t Get the output vector at time t As shown in formula (16):

[0031] Through the above method, the strong nonlinear representation ability of the KAN network is combined with the gated time memory mechanism of LSTM to solve the long-term dependency and nonlinear problems in time series data modeling, and the high-dimensional nonlinear characteristics and long-term dependencies of the current signal can be efficiently extracted.

[0032] Based on the above discussion, in an optional embodiment of the present application, after extracting nonlinear features from the denoised current signal using the KAN network, the process of extracting timing features by the input gated input weight adaptive module may include: In an optional embodiment of S12, the number of KAN network layers is set to 3 when performing high-dimensional nonlinear feature extraction on the denoised current signal.

[0033] S13: Input the time series features into the multi-dimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features.

[0034] It should be noted that the input of time series features into the multi-dimensional feature module mainly includes three channels. The first channel uses a multi-scale network to extract multi-scale features, and the second channel uses spatially separable convolution, which reduces the computational complexity while retaining the full expression capability of spatial features.

[0035] Step 1: Channel 1 divides the time series features into s groups, and each group of features is represented by x i Indicates, i∈{1,2,3…s}, except for feature x1, each x i There is a corresponding (2i+1)×(2i+1) convolution kernel feature, using K i (x i ) represents this operation, y i K i (x i), x1 directly outputs y1= x1 without convolution operation, x2 outputs y2= K2(x2) after 5×5 convolution layer, and other features x i (i ≥ 3) needs to pass through a (2i+1) × (2i+1) convolutional layer and make a residual connection to get the output y i =K i (x i +y i-1 ), the output of each set of features is expressed by formula (17): For all the obtained features, the low-scale features are adjusted to high-scale features by upsampling and then point-by-point addition fusion is performed with the high-scale features. This step is repeated s-1 times to obtain multi-scale features.

[0036] Step 2: Channel 2 uses spatially separable convolution to extract temporal features. In the spatial dimension, the m×m convolution is split into a 1×m convolution and an m×1 convolution. After the spatial convolution, the result of the spatial convolution is convolved with a 1×1 convolution kernel to fuse the channels. The information of multiple channels is merged into a new output channel to obtain the temporal features.

[0037] Step 3: Channel 2 performs maximum pooling of the time series features with a pooling window of z×z and a step size of q to obtain large-scale features. Maximum pooling can retain the maximum value in each pooling area, which helps to retain the key features of high-dimensional variation information.

[0038] Step 4: For the time feature, starting from the leftmost side of the feature vector, take the element of the padding length and arrange the elements in reverse order to obtain the padding vector. Multiply the padding vector and the corresponding position elements of the large-scale feature to generate the corresponding elements of the new vector, and finally obtain the multi-dimensional feature.

[0039] Through the above method, the time series features are input into the multi-dimensional feature module, and multi-scale feature extraction, spatially separable convolution and maximum pooling processing are performed on the time series features through three channels respectively, and finally multi-scale features and multi-dimensional features are obtained.

[0040] Based on the above discussion, in an optional embodiment of the present application, the process of inputting the time series features into the multi-dimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features may include: In an optional embodiment of S13, when channel 1 performs multi-scale feature extraction on time series features, s=4, when channel 2 performs spatially separable convolution on time series features, 1×5 convolution kernels and 5×1 convolution kernels are used, and when channel 3 performs maximum pooling on time series features, a pooling window of 4×4 and a step size of 1 are selected.

[0041] S14: The time series features, multi-scale features and multi-dimensional features are fused using the residual path attention mechanism to obtain the evaluation features.

[0042] It should be noted that for the three types of features obtained above, namely, time series features, multi-scale features and multi-dimensional features, it is very necessary to emphasize the important features and suppress the irrelevant or unimportant features when fusing them. The introduced residual path attention mechanism uses the attention mechanism to suppress unimportant features before feature fusion, so that the fused evaluation features are accurate and effective.

[0043] Step 1: Concatenate the temporal features, multi-scale features, and multi-dimensional features to obtain the input feature map X of the path attention.

[0044] Step 2: Perform average pooling Favg(·) on the input feature map on the channel axis, extract the average features of each path to obtain the path average features, perform maximum and average merging on the path average features, and obtain the representative values ​​vavg and vmax of the time dimension and channel dimension of each convolution path. The i-th element of vavg and vmax can be calculated by equations (18) and (19), respectively: Where K is the number of feature paths, W is the length of the feature, C is the number of output channels, and j and k represent the axes of the time dimension and channel dimension.

[0045] Step 3: Sum the attention scores calculated by the maximum pooling and average pooling to enhance the representation ability of the module and calculate the importance of each path. The vectors of vavg and vmax pass through two FC layers respectively and calculate the path activation vector z, which is calculated by formula (20): Where W1, W2, W3, W4 are the weight vectors of the FC layer, and δ and σ are activation functions.

[0046] Step 4: Perform a scaling operation Fscale(·) on the input feature map X and multiply it with the path activation vector z to obtain the evaluation feature Y, as shown in formula (21):

[0047] Through the above method, while adaptively fusing temporal features, multi-scale features and multi-dimensional features, the useful information of multiple path features is emphasized and irrelevant information is suppressed according to the attention mechanism to obtain accurate evaluation features.

[0048] Based on the above discussion, in an optional embodiment of the present application, the process of fusing the temporal features, multi-scale features and multi-dimensional features using the residual path attention mechanism to obtain the evaluation features may include: In an optional embodiment of S14, the number of feature paths K=3, the activation function δ selects the relu function, and σ selects the SELU function.

[0049] Through the input evaluation features, the deep embedding metric network learns the distance metric between features and obtains a discriminative feature space to judge the state of the sample.

[0050] It should be noted that when classifying based on the deep embedding metric network, it is enough to ensure that samples of the same category are close in the embedding space, while samples of different categories are far away. This intra-class compactness and inter-class separation helps the network make more accurate decisions during classification and improves the accuracy of judgment.

[0051] Input the evaluation features into the linear layer for classification to obtain the prediction vector .

[0052] As shown in formula (22), the mean square error is calculated based on the predicted vector and the actual value y:

[0053] The triplet loss is used, that is, the anchor sample (anchor, ), positive samples (positive, ) and negative samples (negative, ) of the sample group ( ), calculate the ternary loss, where and For similar samples, and By constructing the above triples and inputting them into the deep learning model, we can obtain deep feature embedding in high-dimensional space. . Model optimization will reduce and The intra-class distance of and The inter-class distance. The ternary loss function is shown in formula (23): Where N tr Set the hyperparameter for the number of samples α To avoid vanishing gradients.

[0054] The optimization process of triplet loss is carried out by randomly selecting triplet samples. During the training process, if the selected triplet sample has met the conditions of triplet loss, no operation is required; if the triplet sample does not meet the conditions, the model parameters need to be adjusted to minimize the triplet loss.

[0055] Initialize the anomaly threshold △, use triplet loss to further optimize the anomaly threshold during the training process, adjust the model parameter weights to reduce the model prediction error, and obtain the final anomaly threshold △.

[0056] The state of the sample is judged according to the sum of the mean square error and the ternary loss and the abnormal threshold. If the sum of the mean square error and the ternary loss is greater than the abnormal threshold, it means that the sample is abnormal, otherwise it means that the sample is normal.

[0057] Through the above method, a discriminant space is created and the network is trained through ternary metric error and mean square error to obtain the abnormal threshold. The state of the sample is judged according to the relationship between the sum of the two errors of the calculated sample and the abnormal threshold.

[0058] Based on the above discussion, in an optional embodiment of the present application, for the above-mentioned evaluation features inputted, the distance metric between the features learned by the deep embedding metric network to obtain a discriminative feature space to judge the state of the sample may include: In an optional embodiment of S15, when performing the ternary loss calculation, the hyperparameter α Take 1×10 -3 .

[0059] In summary, the present invention discloses a battery simulator health diagnosis method based on gated input weight adaptation proposed in this application. The method includes proposing a correlation noise reconstruction method to decompose the current signal into multiple sub-signals, and reconstructing the sub-signals using the correlation coefficient method to obtain a denoised current signal; using the KAN network to extract nonlinear features from the denoised current signal, and then inputting the gated input weight adaptation module to extract timing features; inputting the timing features into the multidimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features; using the residual path attention mechanism to fuse the timing features, multi-scale features and multi-dimensional features to obtain evaluation features; through the input evaluation features, the distance metric between the features is deeply embedded in the metric network to obtain a discriminative feature space to judge the state of the sample. The present application provides a battery simulator performance evaluation method, which evaluates the performance of the battery simulator to prevent the poor performance of the battery simulator from causing a decrease in circuit accuracy.

[0060] The present application uses specific examples to illustrate the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the method and core ideas of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A battery simulator health diagnosis method based on gated input weight adaptation, characterized in that: include: The current signal is decomposed into multiple sub-signals using the correlation noise reconstruction method, and the sub-signals are reconstructed using the correlation coefficient method to obtain the denoised current signal; The denoised current signal is input into the gated weight adaptive module to extract the timing characteristics of the signal; After using the KAN network to extract nonlinear features from the denoised current signal, the input gated input weight adaptive module is used to extract the timing features; Input the time series features into the multi-dimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features; The time series features, multi-scale features and multi-dimensional features are fused using the residual path attention mechanism to obtain the evaluation features; Through the input evaluation features, the deep embedding metric network learns the distance metric between features and obtains a discriminative feature space to judge the state of the sample.

2. The battery simulator health diagnosis method based on gated input weight adaptation according to claim 1, characterized in that: It is proposed to use the correlation noise reconstruction method to decompose the current signal into multiple sub-signals, and use the correlation coefficient method to reconstruct the sub-signals to obtain the denoised current signal, including: The correlated noise reconstruction method is used to add specific noise to the collected current signal and then decompose it into sub-signals that can display dimensional feature information. The correlation coefficient method is used to obtain the correlation coefficient between the sub-signal and the original signal. The sub-signals with lower correlation coefficient values ​​are removed, and the weighted sum of the remaining sub-signals is calculated to obtain the denoised current signal.

3. The battery simulator health diagnosis method based on gated input weight adaptation according to claim 1, characterized in that: After using the KAN network to extract nonlinear features from the denoised current signal, the input gated input weight adaptive module extracts timing features, including: The KAN network is used to extract high-dimensional nonlinear features in the denoised current signal. The KAN network can enhance the expression ability, use memory units to store short-term memory and capture the dynamic changes of data in time series, and combine the gating mechanism to dynamically control the information flow to obtain timing characteristics.

4. The battery simulator health diagnosis method based on gated input weight adaptation according to claim 1, characterized in that: The time series features are input into the multi-dimensional feature module for feature extraction and aggregation to obtain multi-scale features and multi-dimensional features, including: The temporal features are input into the multi-dimensional feature module, which mainly includes three channels. The first channel uses a multi-scale network to extract multi-scale features. The second channel uses spatially separable convolution to obtain more delicate temporal features. The third channel uses maximum pooling to retain large-scale features of high-dimensional variation information. The features of the second and third channels are fused to obtain multi-dimensional features.

5. The battery simulator health diagnosis method based on gated input weight adaptation according to claim 1, characterized in that: The time series features, multi-scale features and multi-dimensional features are fused using the residual path attention mechanism to obtain evaluation features, including: The temporal features, multi-scale features and multi-dimensional features are concatenated to generate a path feature map. The original features are compressed in a specific dimensional direction to obtain the representative values ​​of the time and channel dimensions of each path. Then, the attention scores are calculated using the maximum pooling and average pooling respectively and added to obtain the path vector. The path feature map and the path activation vector are multiplied to obtain the evaluation feature.

6. The battery simulator health diagnosis method based on gated input weight adaptation according to claim 1, characterized in that: Through the input evaluation features, the deep embedding metric network learns the distance metric between features and obtains a discriminative feature space to judge the state of the sample, including: The evaluation features are input into the metric classifier to create a discriminant space and the network is trained through ternary metric error and mean square error to obtain the classification threshold. The loss value of the sample in the discriminant space and the classification threshold are used to judge the performance of the circuit.

Images